Generate a custom OSR entrypoint for OSR compiles on all platforms, and transition to optimized cod…

src/arm/builtins-arm.cc

Index: src/arm/builtins-arm.cc
diff --git a/src/arm/builtins-arm.cc b/src/arm/builtins-arm.cc
index eff47e2692bfc2627f4e4269da609b030d39367d..d1ed4d5ac3fd6b7fa0601d7d91a6a4d8bf156ed0 100644
--- a/src/arm/builtins-arm.cc
+++ b/src/arm/builtins-arm.cc
@@ -975,22 +975,35 @@ void Builtins::Generate_OnStackReplacement(MacroAssembler* masm) {
     __ CallRuntime(Runtime::kCompileForOnStackReplacement, 1);
   }
 
-  // If the result was -1 it means that we couldn't optimize the
-  // function. Just return and continue in the unoptimized version.
+  // If the code object is null, just return to the unoptimized code.
   Label skip;
-  __ cmp(r0, Operand(Smi::FromInt(-1)));
+  __ cmp(r0, Operand(Smi::FromInt(0)));
   __ b(ne, &skip);
   __ Ret();
 
   __ bind(&skip);
-  // Untag the AST id and push it on the stack.
-  __ SmiUntag(r0);
-  __ push(r0);
-
-  // Generate the code for doing the frame-to-frame translation using
-  // the deoptimizer infrastructure.
-  Deoptimizer::EntryGenerator generator(masm, Deoptimizer::OSR);
-  generator.Generate();
+
+  // Load deoptimization data from the code object.
+  // <deopt_data> = <code>[#deoptimization_data_offset]
+  __ ldr(r1, MemOperand(r0, Code::kDeoptimizationDataOffset - kHeapObjectTag));
+
+  // Load the OSR entrypoint offset from the deoptimization data.
+  // <osr_offset> = <deopt_data>[#header_size + #osr_pc_offset]
+  __ ldr(r1, MemOperand(r1, FixedArray::kHeaderSize +
+      DeoptimizationInputData::kOsrPcOffsetIndex * kPointerSize
+      - kHeapObjectTag));
+  __ SmiUntag(r1);
+
+  // Compute the target address = code_obj + header_size + osr_offset
+  // <entry_addr> = <code_obj> + #header_size + <osr_offset>
+  __ add(r0, r0, r1);

[vincent.belliard.fr, 2013/07/31 15:07:44]

You can do the SmiUntag and the add in one instruction:
add(r0, r0, Operand::SmiUntag(r1));

+  __ add(r0, r0, Operand(Code::kHeaderSize - kHeapObjectTag));

[vincent.belliard.fr, 2013/07/31 15:07:44]

the result of add can be directly stored in lr. It avoids the mov.

+
+  // Overwrite the return address in the link register.
+  __ mov(lr, r0);
+
+  // And "return" to the OSR entry point of the function.
+  __ Ret();
 }